FR2670923A1 - CORRELATION DEVICE. - Google Patents

Abstract

Correlation device calculating a correlation between a current block of a first image and blocks of symbols taken from a second image to determine displacement vectors. Selection means (26) (27) select, in a limited number, symbols contained in said blocks on which the correlation calculation relates. This selection can be made either by random selection or as a function of the informational content of the blocks. Said current block can be subdivided into zones on which said selection means operate. A correlator (25) performs the block-to-block correlation calculations over the entire current block and a decision unit (28) determines the displacement vector. It is possible to choose for the correlator (25) an architecture which favors either its computation speed or a reduction in its hardware complexity. The invention finds its application in reducing the data rate to be transmitted, in particular in image processing (television, videophone) or sound processing. Application: Television, videophone, sound processing.

Description

DESCRIPTION

"CORRELATION DEVICE".

  The invention relates to a correlation device

  operating on input signals representing symbols

  in blocks in a regular configuration, including

  nant: a correlator which determines a correlation between symbols of a current block and symbols having identical positions in a sequence of blocks to be processed from said configuration, and selection means which select among the symbols of said current block a number limited symbols of

  reference to which the correlation relates.

  Regular configuration of input signals

  can be one-dimensional or two-dimensional It can be

  for example a one-dimensional configuration o

  correlations can occur on blocks of

  sound source for sound processing A symbol

  the configuration is then a value coded in binary form Preferentially it can be configurations

  two-dimensional solutions where correlations take place on

  eg, animated, on which we determine

  movers of the animated sequence between successive images A symbol of the configuration is then a pixel These applications are often undertaken with a view to a reduction of the data rate.

  in the field of television (high speed) or in the field of

  videophone and video conference (low speed).

  The application relating to the estimation of movement

  television has the particularity of revealing very severe constraints, particularly

  need to operate a large amount of calculations with a

  high bit This application will be taken here as a

  example without this limiting the invention.

  For motion estimation motion, the technique that seems the most interesting is the so-called "block matching" called "block matching" in English It consists of cutting the image into blocks of pixels For a given block of the image is determined in an earlier image the block that comes closest to determine the movement of the animated sequence shown in

  image It is this movement information that is transmitted

  for each block, to the receiving member placed at the end

mity of the transmission channel.

  The principle of treatment consists of following the

  evolution of the content of images on successive images This technique takes advantage of the fact that the information content

  varies little from one image to another The movement of information

  This monitoring is done by identifying certain parts that may be the most significant or that are chosen voluntarily. This may concern all the blocks of

the image.

  A current reference block (simplified expression

  subsequently "current block") may be located

  differently in either anterior or posterior images

  It is this displacement that needs to be determined. But this displacement remains limited and for this we define a window within which the current block is likely to

be located.

  The estimated movement is the translation between the

  position in the image of the current block and the position in the i-

  anterior / posterior block of the anterior / posterior block with the strongest correlation

  a vector of displacement relative to the current block The correlation

  tion will be determined pixel by pixel of the same position

  for all possible positions of the current block in the

  in the corresponding window The same operations are

  repeated for the other selected current blocks The correlation

  tion is summed for the whole block, the absolute values of the differences in

  pixel to pixel block to block The characteristic of the pixel is large

usually its luminance.

  Two main obligations appear for a material realization of a motion estimator: its

  ability to operate in real time, its material complexity.

  Different strategies have been developed, but some

  Gorithms that are less demanding in number of operations are less regular and therefore more complex to implement The regularity of the calculation phases is a condition that must be taken into account in particular for the realization of

  Dedicated integrated circuits This explains that the usual strategy

  implementation is a comprehensive search

  covering all the pixels of all the blocks of the window.

  Scanning techniques for successively operating the analy-

  each pixel of the current block in the window can

  to reduce the material complexity of the implementa-

  tions. We therefore sought to develop methods to reduce the complexity of motion estimation processing without impairing its ability to operate in real time.

  knows the article entitled "Motion-compensated interframe co-

  ding for video conferencing "by T KOGA and National Telecommunications Conference 29 November-3 December 1981

  NEW ORLEANS LOUISIANE This document defines a vector of

  movement by involving different types of

  especially a subsampling that allows

  take into account that a limited number of pixels per

  rant that is to say a pixel on two in a line for a first case and which extends even to a line on two in a second case The hardware complexity is diminished but at the expense of the accuracy of the motion estimation

  inter-image differential coding, which is a primary application

  the motion estimation, consists in transmitting a prediction error, the latter being in the case

  present the block of the previous image chosen in the window.

  When the motion estimation is done correctly the prediction error is small and therefore the length of the code transmitted by a transmission channel is reduced. It is therefore conceivable that the degradation of the accuracy of the motion estimation has an impact on the encoding efficiency Thus in the cited prior art for a transmission rate of 1.5 Mbits / s the rate of prediction errors can increase by 50% This requires an increase in bit rate for

  achieve acceptable image quality.

  The problem is, therefore, to devise a

  correlation which allows a simplified material realization (compared to a usual realization) and which

  can operate in real time with a predictive error rate

  reduced even for image processing applications

of TV.

  This goal is achieved with a correlation device

  whose selection means select, in said current block, a reference symbol (s) by choosing its (their) position: either according to the informational content of said current block, or according to a random selection, or for a first part primarily according to the informational content and for a second part according to a

random selection.

  It is thus possible to adopt two strategies

  complementary that can be combined.

  In the case of an image for example, advantageously

  According to the first strategy only the pixels carrying "distortion" information in the current block can

  be chosen for the correlation calculation.

  According to the second strategy, we ignore the content

  information of the current block, and the reference pixels se-

  selected are randomly distributed in the current block.

  These two strategies make it possible to break the

  of the selection that is made when eliminating a pi-

  xel out of two or even a line on two The regularity of this selection has the disadvantage of eliminating

  the high frequency part of the information and

  However, according to the invention, the high frequency portion of the information spectrum is not suppressed and its importance can even be controlled by intervening on the limited number. of symbols selected by current block By combining the strategy by the informational content and the strategy by random selection, this absence of regularity of the position of the symbols in the current block is preserved and the error rate is further reduced by following closely the content of the information In the latter case the reference symbols are selected primarily according to the informational content and their number is completed by a selection

  random, up to the predetermined limited number.

  When a block consists of pixels (

  binary coded values) the selection consists in taking reference of pixels (respectively values coded in binary) having a certain position in the

current block.

  To these pixels that have a determined position in the current block correspond other pixels that have the same

  position in blocks of the window to be analyzed The correlation

  which is calculated between two blocks is actually between

  the pixels of the two blocks that have the same position.

  continued for all possible positions of the

  current block in the window The treatment is then

  followed by another current block associated with another search window It is possible for the same random selection to be used for several current blocks to select the position of the selected reference symbol (s) randomly. In the case where the selection is made taking into account the information content of the current block, the position of the reference symbols in the current block is determined from a distortion function calculated for each symbol of the current block vis-à- vis the surrounding symbols, the selected reference symbol (s) being chosen

  among the symbols whose values of the distortion function

  most different from those concerning the environmental symbols.

  ronnants. This distortion function can, for example, take into account the variation of a characteristic such as luminance or others between the neighboring pixels of the same current block To determine the most representative pixels

  of a current block, the distortion function can be deter-

  by calculating a Laplacian pertaining to the characteris-

  tick chosen for the input signal More generally

  can take into account neighboring pixels that surround a

  given in an isotropic way But in order that in a

  reference pixels are not concentrated in one

  too localized part of the current block, it is possible to

  a distribution constraint by subdividing the block

  in several areas, for example in the form of a quadrilla-

  regular or not regular or any other form of mesh In this case it is required that all zones or at least part are represented by at least one pixel This pixel per zone

  can be chosen with the same strategies as those already de-

  For example, when a current block is subdivided into several zones, the selection means operate separately on each zone, the correlation always taking place.

  on all the reference pixels of the current block Pre-

  ferentially each zone of a current block has a same

  number of reference pixels When the pixel (s) of

  zones are (are) selected from the

  termination of a distortion function, preferably the reference pixel (s) is (are) selected in each zone so that the values of the distortion function which characterize it (s) are alternately

  maximum and minimum for contiguous areas.

  According to the invention, it is not stored in a block

  only a limited number of reference symbols for

  Correlation calculation This limited number can be chosen in several ways It can be fixed a priori after preliminary tests which allow to establish a correlation calculation with an acceptable error rate. This can be done when the current block contains a reduced number

pixels for example 4 x 4.

  In the case of a strategy implementing a distortion function that takes into account information content, this limited number can be chosen from limits imposed on the values that this function can present.

distortion.

  More generally, when the size of the block

  becomes large, it is possible to take into account the

  informational nude of the entire image and not just

  of the current block for the determination of the

  For this purpose, a correlation matrix is established.

  characteristic of all the input signals con-

  By the so-called "principal component analysis" method, the eigenvalues of this correlation matrix are calculated and the corresponding variances are determined among which the number of highest variances necessary for a good representation of the entire image This number is then chosen as the limited number of

reference symbols.

  According to the invention, in this case, the limited number of reference symbols is thus dependent on the content of the image. In a large number of situations, this limited number remains mainly related to the type of images, that is to say to the application itself Nevertheless it is possible to enslave more finely said limited number to the contents of the image by determining it regularly and adaptively on batches of images This can be obtained using a device of correlation which comprises a neural network which implements the so-called 'principal component analysis method' and which selects the number of the highest variances Cyclically this number of variances is introduced into the selection means to constitute said

  limited number to operate as already described.

  A neural network is a known device of

  signal processing which can, after a programming step,

  or a learning step, perform treatments

  for example, a principal component analysis

  Such an array of neurons can for example be the one described in the document

  Optimal Unsupervised Learning in Single-Layer Linear Feed-

  forward Neural networkso T D SANGER Neural Networks, Vol 2,

1989, p 459-473.

  As indicated above, the fact of

  the correlation calculation should only be applied to certain sym-

  reference boles having a certain position in the block

  current makes it possible to simplify the material realization,

  to increase the speed of treatment while keeping the rate

  from errors to low and controllable values.

  In addition, the mechanisms for selecting reference symbols make it possible to provide in certain situations

  further simplification of the complexity of the

  In particular, when the current block is subdivided into zones, this advantageously makes it possible to obtain

  either a gain in processing speed or a gain in comple-

  This situation appears when the

  tif correlation operates a correlation calculation with

  current blocks of data subdivided into p areas with n / p sym-

  reference bots selected by zone, the correlator

  taking: a systolic array of registers which transfer to one another the data of the pixels of the window, a plurality of processors, each processor performing a computation of the correlation of at least one displacement vector, the setting device implement a first calculation step during which:

  a) each processor performs a successive correlation calculation

  on the n / p selected reference symbols of a first zone, said calculation successively bearing on all the motion vectors that said processor must calculate,

  b) storage means store results between

  issued by the processors after a),

  (c) transfer means transfer the contents of the

  authorities operating in one area to subsequent registers operated

  rant on a next zone, said first calculation step then being repeated for the following p-1 zones,

  means of summation determining the sum of the results

  intermediaries relating to each block to be processed to deliver

  the correlations of said series of blocks.

  A particularly interesting case arises for the image processing, when said zones contain several pixels of which only one is selected to perform the correlation calculation It is then possible to obtain an additional gain and to choose, for the correlator, between a architecture that favors speed or architecture that favors the simplicity of hardware that is to say that reduces

  the surface for an integrated realization.

  In the case of architecture emphasizing

  the correlator comprises substantially as many processes

  that there are possible displacements in the window and it operates on a reduced number of reference pixels which increases its speed of execution compared to a use

  of all the pixels in the current block A special case

  Interestingly, where each zone contains only 4

  pixels, with only one reference pixel selected per zone.

  In this case, the correlator operates on a two-dimensional window.

  pixels with common blocks subdivided into zones

  containing 4 pixels each with a single reference pixel selected

  selected by zone, a two-dimensional matrix of registers

  storing, at a given moment, the pixels of said window se-

  the same scheduling, four registers being assigned to the

  treatment of an area, rows of said registers

  between them, in the two directions of the matrix, pixels of even order of the window and other rows of registers communicating to each other odd order pixels of the window, this communication being unidirectional in a direction of the window and being bidirectional in the other direction of the window, a processor being assigned to the processing of each displacement vector and operating with four registers containing even and odd pixels of the window, the order numbers of said pixels taken two to two being consecutive in each direction, a common selection signal causing the processors to operate with the data of the registers corresponding to the reference pixels

  selected in the current block.

  In the case of the architecture favoring

  complexity of the hardware, a particularly interesting case is where the correlator operates on a two-dimensional pixel window with common blocks subdivided into rectangular zones containing k pixels with a single pixel

  selected by zone, a two-dimensional matrix

  the registers storing, at a given instant, the pixels of said window according to the same scheduling, k registers being

  assigned to the treatment of a zone, rows of said registers

  very two-way communication between them pixels

  of the window in both directions of the matrix, a pro-

  the processor being assigned to the successive processing of k displacement vectors by operating with k registers containing pixels belonging to the same zone, the order numbers of said pixels

  being consecutive in both directions, a se-

  making each processor operate for each

  with the data of the register corresponding to the pixel of

  selected reference of the current zone.

  The invention is not limited to the case of configura-

  two-dimensional representations of symbols It also applies

  in the case of one-dimensional configurations The symbols can

  The correlation device then operates on current blocks containing a plurality of said values from which are selected, in limited number, and in position

reference.

  The invention will be better understood with the aid of the following figures given as non-limiting examples which represent:

  figures IA, l B: a representation of an image in the-

  which is a window containing blocks to be treated on which the correlation must relate and that of a current block in a posterior image, FIG. 2: a block diagram of a correlation device, FIGS. 3A, 3B: a representation of a current block of 8 × 8 pixels indicating the positions of the reference symbols selected according to the prior art, FIGS. 3 C, 3 D, 3 E, 3 F: by way of example, a representation of a block current of 8 x 8 pixels indicating the positions of the selected reference symbols according to the invention,

  FIGS. 4A, 4B, 4C, 4D: a representation of the pixels

  a given pixel for calculating the function of dis-

  torsion, figure 5: a diagram of some of the means of selection

  to calculate a distortion function according to the

presentation of Figure 4A,

  Figure 6: a detailed diagram of a joint selection unit

  taking a calculating member of a Laplacian and a sorting member, FIG. 7: a block diagram of a correlation device comprising a neural network for determining the number of reference symbols,

  FIG. 8: a diagram of a correlator according to the prior art

  9 A, 9 B: a representation of blocks to be treated of a window and a current block divided into zones, FIG. 10: a diagram of a processor surrounded by the basic elements necessary for its operation,

  Figure 11: a diagram of a correlator favoring simplicity

  plicity of the hardware operating on current blocks subdivided into zones of 4 symbols among which is chosen a reference symbol.

  figure 12: a diagram of a correlator favoring the

  working on current blocks subdivided into zones of 4 symbols. FIG. 1A represents an image 10 and FIG.

  1 B represents an image which may be anterior but

  posteriorly to image 10 on which is

  performed the correlation calculation.

  The image 20 is formed of pixels that are analyzed in the form of current blocks, for example the blocks 211, 212, 213 in the image 20, for any current block by

  example, the current block 22, the processing will consist of de-

  finish at which block of the previous image 10 it is correlated.

  For this purpose around a block 12 which has, in the image 10, the same coordinates as those of the current block 22 in the image 20, a window 13 is determined which surrounds the block 12 in order to limit the number and the position blocks to be searched as most likely to contain the correlated block The size and shape of this window depends on the extent of the search that is desired and therefore

  definitive treatment time that it is possible to

  create a window and also the complexity that

  we admit for the material realization that follows from it.

  Thus the current block 22 (image 20) is compared to all the blocks of the window 13 of the image 10 obtained by moving the

  block in successive steps of at least 1 pixel.

  At the end of this processing another current block is selected from the image to which another window in the image will correspond and the same processing is performed, and so on. The current blocks selected in the image 20 can be to be limited in number but preferentially all

  the blocks undergo the same treatment Depending on the

  of the current block in the image 20 and that of the block correlated in the image 10, a vector is deduced each time.

  of the current block This is the vector of

  as well as the pixel feature deviation (eg

  luminance) which is encoded and transmitted through the

  transmission in order to operate data compression.

  The correlation device which implements the

  treatment is shown in Figure 2 It includes a cor-

  25, a selection unit 26 which determines on which

  pixels of the current block 22 must carry the correlation, a uniform

scanning station 27 which determines the order in which the pixels of the block 12 and by extension of the window must be analyzed at the output of the correlator 25 a decision unit

  28 determines the block of the window 13 for which the correlation

  The current block BL to be processed is introduced in the selection unit 26 and in the correlator 25. The latter also receives the range of the window FIRE which can

  to be slaved to the speed of execution necessary for the application

  tion. The correlation consists in successively comparing two blocks with each other by comparing the pixels having in each block respective identical positions. It is common practice to perform an exhaustive processing involving all the pixels of the blocks. However, according to the document already cited, it is possible to reduce the importance of treatment in

  whereas a limited number of pixels per current block.

  Figures 3A and 3B show two possibilities.

  These are a current block of 8 x 8 = 64 pixels on which the reference pixels which have been selected to perform the correlation are hatched. In FIG. 3A, one pixel out of two is selected. divides the processing by 2 In Figure 3 B is also ignored a complete line of pixels in two which further divides the treatment by 2 These limitations of the number of reference pixels can be easily done for example by intervening on the durations of But the great regularity that results from these modes of limitation has the consequence that the high frequency part of the spectrum of a

  block of information is truncated and an important error rate

so much appears.

  To remedy this defect according to the invention, systematic selection of the

  reference pixels within a current block.

  A first strategy shown in FIG. 3C as an example consists of a random selection up to a predetermined limited number of reference pixels. Thus, in FIG.

  OS reference selected randomly in the current block.

  A second strategy shown in FIG. 3E by way of example takes into account the information content contained in the current block. Thus, suppose that a fictitious curve 31 delimits in the current block two different luminance portions according to the invention. selected to perform the correlation will be determined taking into account the information content by determining a distortion function Briefly we can indicate that the selection will take place in the parts where the characteristic (here the luminance) undergoes variations with respect to close neighbors All the pixels of the current block are thus examined The selection unit 26 (FIG. 2) makes this determination FIGS. 4A, 4B, 4C, 4D indicate some possible modes of examination of the informational content

with isotropic distributions.

  FIG. 4A represents a preferential mode which consists for a given pixel (hatched) to compare its luminance variation with its neighbors having coordinates individually distant from 1 This amounts to calculating a Laplacian A such that: A = 4 I (x, y) - (I (x + 1, y) + I (x, y + 1) + I (x-1), y) + I (x, y-1)) For Figure 4B the distortion function is : r = 4 I (x, y) - (I (x + 1, y + 1) + I (x + l, y-1) + I (x-1, y + 1) + I (x-1 l, y -1)) For Figure 4 C the distortion function is: A = 4 I (x, y) - (I (x + 2, y) + I (x-2, y) + I (x, y + 2) + I (x, y-2)) For Figure 4 E the distortion function is: Y = 8 I (x, y) - (I (x + 1, y + 2) + I (x + 1, y-2) + I (x-1, y + 2) + I (xl, y-2)

  + I (x + 2, y + 1) + I (x-2, y + 1) + I (x + 2, y-1) + I (x-2, y-1)).

  Other distributions of surrounding pixels

  are possible even if they are not isotropic.

  When the distortion function has thus been calculated for all the pixels of the current block, the selection unit determines the pixels having the extreme values

  to select them to represent said current block.

  According to the two strategies which have just been

  written it appears that some parts of a current block

  may not be represented or not represented.

  To limit the error on the correlation which risks to re-

  it may be desirable for those parties to be

  less partially represented For this purpose more

  According to the invention, the current block is subdivided into zones

  which may be regular or not preferentially

  divides the current block into identical areas with the same number of pixels Then the random selection is made

  not at the level of the current block but at the level of the zone.

  Thus the reference pixels are selected randomly by zone This is represented in FIG. 3 D o 16 randomly selected reference pixels are represented by hatches The same random selection of this type can be used for several current blocks to operate the calculation

  correlation Although the random selection takes place at

  zone, the correlation calculation shall be carried out at

current block level.

  Figure 3F represents the situation where both

  strategies can be combined For this a first part

  of the reference pixels is selected in priority according to the informational content and a second part

  is selected from a random selection by zone.

  This second part is limited to the extent that

  all of the selected reference pixels reaches the name

limited number.

  The diagram of FIG. 5 represents a selection unit 26 determining the position of the points in a block

  current from the calculation of a distortion function (function

  A, r, A or Y preceding) performed in a calculating unit The results obtained are sorted in a sorting organ 51

  according to a sorting algorithm in order to limit the number of symbols

  the current block references The addresses of the selected reference symbols are stored in an address register

  52 to be operated by the correlator The selection unit

  tion 26 operates under the control of the scanning unit 27.

  A detailed diagram of the calculating member 50 and the sorting member 51 is shown in FIG. 6 in the case

  o the distortion function is the Laplacian A defined previously

  This distortion function corresponds to the

  of Figure 4 A where the letters N, S, E, W represent

  The letters are shown in the diagram of figure 6, the letter C representing the central pixel. The situation concerns the case where the scanning applies to all the pixels of the

  current block according to line-by-line scanning

  6 is a chain of delays consisting of: a first delay line 60 and a second delay line 63 which each delay the data of the duration of a line minus that of a pixel, a first register 61 and a second register 62 which each delay the data of the scanning duration of a

pixel.

  So after three scan lines, the data

  pixels N, S, E, W are distributed as is

  shown in FIG. 6: the datum N is at the output of the delay line 63 and

  the data S is at the input of the delay line 60.

  An adder 64 and an adder 65 operate respectively additions S + E and W + N The two results are added in an adder 66 This last result must be subtracted by 4 times the value of the central data C This is in C between the two registers 61 and 62 To multiply it by 4 it suffices to shift it by 2 bits

  at the input of the subtractor 67 which receives the output of the addi-

  66 The output of the subtractor 67 delivers the function

  distortion A which is calculated for all the pixels of each

  that current block Among all these pixels we select a limited number to form the reference pixels using the sorting member 51 According to a preferred mode is sorted in the set of all the pixels of a current block the pixels for which the distortion function is either maximum

  is minimal and this sorting is applied to the case of

  by having two contiguous zones possess func-

  A alternating positive and negative.

  For this purpose the sorting member 51 comprises a sign selector 70 which is controlled by the scanning unit 27, a battery of registers 71 storing maximum / minimum values for each zone,

  a comparator 72 which determines for each zone the value

  keep them either maximum or minimum according to the required sign. The addresses of the selected reference pixels are stored in the address register 52 which receives the ADP pixel addresses and the ADZ zone addresses of the unit.

sweep 27.

  The scanning unit 27 gives the rate of signs

  alternates + and assigned to each zone When the sign po-

  is assigned to a zone, the result from the calculating organ 50 is taken with its sign, then compared to the value

  previous stored for the same area in the battery pack

  71 If the last value is the highest, it replaces

  place the old value in the battery of registers 71 Si-

  simultaneously the address of the corresponding pixel ADP is stored

in the address register 52.

  When the negative sign is assigned to another zone, the result coming from the computing unit 50 is taken by inverting its sign. It is compared with the previous value stored for the same other zone in the register bank 71. sign inversion, the same comparator operates in the same way in both cases by determining a maximum. At the end of the scanning, the address register 52 contains the sorting result, that is to say, for each zone, the address of the selected pixel, having the maximum distortion

or minimum of the area.

  The diagram of Figure 6 corresponds to the calculation of

  the distortion function relating to Figure 4 A

  as for the same method, the person skilled in the art can easily establish the diagram of the calculation unit operating the calculation of the examples of distortion functions of FIGS. 4B, 4C and 4D. To carry out a compact and regular implementation according to this aspect of the invention it is interesting that the correlation device operates on blocks of Nx N pixels, with N = 16, of which are selected n

reference pixels, with n = 8.

  Until now the number of reference symbols in a current block was limited and predetermined in advance, the selection taking place on the position of the selected pixels in the current block. The number itself can be determined in several ways. the number of pixels per current block is small, it is possible to carry out preliminary tests according to the type of images (and therefore of the configuration) to be processed. The number of reference pixels selected is reduced at each test. function of errors between the original image and the restored image and this number is determined according to the acceptable level of errors The same number can be applied to

several types of images.

  When the number of reference pixels per block

  the method can be increased by preliminary tests.

  It is therefore preferable to analyze the images

  characteristics that are representative of the distribution of the pixels In this case the invention recommends

  to use the so-called "principal component analysis" technique.

  blades "of said distribution (taken here in the mathematical sense

  term) The eigenvalues of the correlation matrix relative to this distribution are calculated and the corresponding variances from which the number of

  highest variances sufficient to represent

  It is this number of variances that is chosen as a limited number of reference pixels. This calculation can be done beforehand for standard images and then exploited in the correlation device. But it is also possible to adapt it on the images provided to the correlation device Thus dynamically the limited number of reference pixels can be adapted to the images to be processed in order to operate a more precise treatment This calculation is operated at a rate that depends on the images to be processed by a neural network that uses the "principal component analysis" technique This is represented in the diagram of FIG. 7 which has the same elements as FIG. 2, in addition to a neural network 29 which receives the image blocks to be processed and which provides the selection unit 26 the

  limited number of reference pixels to select.

  Figure 8 shows the schematic of a structure

  classical correlator It consists of systematic networks

  cells of a b cells where a and b are the dimensions of the

  Each cell contains a register R which

  memorizes the value of a sample and an elemen-

  P, connected to the register R, which calculates the function of

  The correlator contains as many cells as the window contains pixels to analyze Each cell has a specific position in relation to the window, so a

  displacement vector of its own.

  Figure 8 shows a 9-cell correlator

  A first rank contains the cells Rjj / Pji to R 13 / P 13, a second rank contains the cells R 21 / P 21 to R 23 / P 23, a third rank contains the cells R 31 / P 31 to R 33 / P 33 Each register is connected to its four closest neighbors by 3

  receiving links and 3 transmitting links. For example,

  the register R 22 receives data from the registers R 12, R 21, R 32 The same register R 22 delivers data to the registers R 12, R 23, R 32 Each register contains successively the data concerning all the pixels of a block The processor assigned to said register calculates each time the correlation function between this data item and the ECU value of

  the current sample which is distributed on all the processes

  If c and d are the dimensions of the block to be analyzed, it is necessary to cycle each processor to accumulate the total of the distortion function relative to a block to be analyzed. At each cycle, all the processors are distributed with the value of 1. sample of the current block and the set of contents of the registers R are shifted by one step so that for every processor there is always the same displacement D between the

  data of the register R and the data disseminated

  xes (not shown) can determine the processor delivering the best correlation and deduce the displacement vector by the position of this best processor in

  the network It is therefore necessary to have cells operating on cycles.

  In the structure of FIG. 8, a processor is assigned to a register for processing all the pixels of a current block. According to the invention, this number of pixels to be processed is limited.

  get a gain on either speed for a hardwa-

  is substantially similar, either on the size of the hardware for

a substantially constant speed.

  Substantial advantages are obtained when the current block is subdivided into zones. FIGS. 9A, 9B show schematically the processing that is carried out. FIG. 9B shows a current block 22 which is subdivided into 6 zones, for example the zone 221. same pixels A zone contains 9 pixels in this example among

  which are selected the reference pixels.

  Moreover, in an image 10, is represented

  a window to be treated 13 The treatment of the totality of the

  mage consists of performing successive treatments on a

  following such windows As has been described previously

  Correlation computation consists of correlating pixel to pi

  xel the current block 22 with each block to be treated constituting the window, for example the block 12 All the blocks to be treated included in the window are obtained by sliding a step of 1 pixel successively in both directions The block 12 slides of a pixel successively according to the block 121 (points) then according to the block 122 (cross) They are slightly offset vertically to be able to represent them In the case where the current block 22 is subdivided into zones, the calculation of

  Correlation with the blocks to be processed takes place at two levels.

  We first consider the reference pixels of a given zone, for example those of the zone 221 of the block

current (top left).

  The correlation is calculated for the pixels of each

  that block to be processed (in a window) having the same relative position in the block, that is to say for the same area in the upper left corner in the block. Thus, intermediate results are obtained concerning a particular area for all the blocks. The calculation then relates to the next zone according to

same process.

  The treatment is schematized in the following way

  te: first zone first block second block second zone first block second block then following zones then following windows

  The progress of the calculations in this order

  the benefit of reducing data exchange between

  This simplifies the architecture and reduces the calculation time. Figure 10 shows the basic components needed to perform the correlation calculations The diagram represents only one processor and one register but the situation is the same for the other processors and the

other registers.

  Any register 40 is connected to other neighboring registers for example in the four directions

  N, S, E, W by bidirectional or unidirectional buses se-

  In the architecture used These data transfers between the registers are controlled by control means 71 common to all the registers A processor 72 is connected to several of these registers by connections Il, I 2, I 3

  which transmit the data of the registers.

  This number of connections also depends on the

  chitecture used The processor 72 operating the calculations in successive steps must temporarily store intermediate results relating to the zones, themselves subdivided according to the blocks These intermediate results are stored in

storage means 73.

  When all the intermediate results have been obtained, they are summed in summation means 74

  to deliver for each block the totality of the results

  termediaries relating to all areas of the block On

  obtains a series of correlation values that relate to

  to the different blocks from which is selected the one which, having the best correlation, constitutes the block whose position constitutes the new position of the current block in the image. Figure II shows the diagram of a part

  of a correlator according to the invention which favors a reduction

  In this case, we no longer associate a processor P with a single vector of displacement but we have it operate for several vectors of displacement for example f

  displacements (thus f registers) In this case there is a b regis-

  very R and a b / f processors When this value f is equal to the number of points per zone, ie f = c d / n, we can take advantage of some of the arrangements already presented. In this case each register is connected to its nearest neighbor registers.

  and any processor (e.g. P 81) receives the data

  generated in the surrounding registers (respectively R 81, R 82, Rg 1, R 92) With for example zones of 4 pixels

  (FIG. 11) from which one reference pixel is selected

  this, to process a window of 4 x 4 pixels then 16 registers and 4 processors All processors receive a selection signal SEL which indicates the reference pixel selected in the area This architecture is easily generalized to the case of areas having k pixels among which one selects a single reference pixel per zone Each processor delivers the results of its calculation on a proper output, for example 581 for the processor P 81 Plus generally when a and b are the dimensions of the window, c and d the dimensions of the current block with N reference pixels selected by area the correlator has ab

  registers, has b / f processors and operates on c d cycles.

  FIG. 12 represents the diagram of a repetitive part of a correlator according to the invention which favors

  the speed gain in the case of a division of the

  Zone diagram This diagram corresponds to zones of 4 pixels from which only one reference pixel is selected to perform the correlation. The processing that is to be performed for a particular zone will be performed by the organs located in the zone.

  inside the discontinuous reference mark 90 There are four registers

  very R 51, R 52, Rs1, Rs 2 all connected to the same processor P 51.

  Each register is no longer connected directly to the register that precedes it and to the one that follows it as in Figure 8 but the connections in both rows and columns jump one

  register in two This provision stems from the fact that

  selecting a single pixel on 4 pixels per zone we must jump

  (Figure 3 D) one pixel out of two in both directions.

  A selection signal SEL makes it possible to operate the processor on one of the four registers which surround it in which are stored at a given moment the values of the 4 pixels of the zone. In a more general case when the block contains cd pixels among which one selects N reference pixels so it is necessary cd / n connections between each processor and the neighboring registers The ECH value of the sample

  is simultaneously broadcast to all processors for calculating

  ler correlation With a window of a b blocks, it takes a

  correlator of a b cells operating on N cycles each

  the seller delivers the result of his calculation on a

  pre, for example 551 for the processor P 51.

Claims (12)

  A correlation device operating on input signals representing symbols organized into blocks in a regular pattern, comprising: a correlator which determines a correlation between symbols of a current block and symbols having identical positions in a sequence of blocks to be processed from said configuration, and selection means which select from among the symbols of said current block a limited number of reference symbols to which the correlation relates, characterized in that the selection means select, in said current block, a reference symbol (s) by choosing its (their) position: either according to the informational content of said current block, or according to a random selection, or for a first part primarily according to the informational content and for a second part according to a random selection.   2 Device according to claim 1 characterized   in that the same random selection is used for several   current blocks to select the position (s)   reference symbol (s) selected randomly.   3 Device according to claim 1 characterized in that said position adapted to the informational content is determined from a distortion function calculated for each symbol of the current block with respect to the surrounding symbols, the symbol (s) of selected reference (s) being selected from the symbols of the current block whose values of the distortion function differ the most from   those relating to the surrounding symbols.   4 Device according to claim 3 characterized   in that said distortion function is determined by cal-   a Laplacian referring to a characteristic of his own to the input signal.   Device according to one of claims 1 to 4,   characterized in that the current block is subdivided into several zones, the selection means operating separately on each zone, the correlation taking place on all the symbols of   selected reference from the current block.   6 Device according to claim 5 characterized in that the number of reference symbols is the same for each zone.   Device according to claims 5 or 6,   in that the reference symbol (s) of an area   is (are) selected so that the distortion function   alternatively maximum and minimum for contiguous areas.   8 Device according to one of claims 1 to 7   characterized in that it comprises a neural network which de-   adaptively terminates the number of reference symbols according to the informational content by calculating eigenvalues of the correlation matrix and the variances, said number of symbols being equal to the number of most variances.   high characteristics of the set of input signals.   9 Device according to one of claims 1 to 8   characterized in that symbols organized according to a configuration   regular ration are pixels constituting an image, the correlation device operating on blocks of Nx N pixels   reference numbers from which are selected N pixels of   ference (n <N 2), the correlations being determined between a current block and blocks of a window taken in the image, the current block being derived from anterior or posterior image   to determine the motion vectors of the blocks of the i- mage.   Device according to Claim 9, characterized   in that N = 8 pixels and n = 16 pixels.   11 Device according to one of claims 1 to 8   characterized in that the symbols are bi-coded values   representing a one-dimensional configuration of   acoustic signals, the correlation device operating on current blocks containing a plurality of said values among   which are selected, in limited numbers and in reference values.   Device according to one of claims 1 to 11   which calculates the displacement vectors relative to a sequence of data blocks to be processed contained in a window, characterized in that the correlator performs a correlation calculation with current data blocks subdivided into p zones with n / p symbols of selected by zone, the correlator comprising: a systolic array of registers which transfer to one another the data of the pixels of the window, a plurality of processors, each processor performing a calculation of the correlation of at least one displacement vector, the device implementing a first calculation step during which:   a) each processor performs a successive correlation calculation   the selected n / p reference symbols of a first zone, said calculation successively   the displacement vectors to be calculated by the said sor,   b) storage means store results between   mediators delivered by the processors at the end of a), c) transfer means transfer the contents of the   registers operating in one area to subsequent registers where   on a next zone, said first calculation step then being repeated for the following p-1 zones,   means of summation determining the sum of the results   intermediaries relating to each block to be processed to deliver   the correlations of said series of blocks.   13 Device according to claim 12, characterized   in that the correlator operates on a two-dimensional window   pixels with common blocks subdivided into recom-   tangulars containing k pixels with a single reference pixel   selected by zone, a two-dimensional matrix of regis-   very storing, at a given instant, the pixels of said window according to the same scheduling, k registers being assigned to the processing of a zone, rows of said registers communicating with each other bi-directionally of the pixels of the window in the two directions of the matrix , a processor being assigned to the successive processing of k displacement vectors by operating with k registers containing pixels belonging to the same zone, the order numbers of said pixels being consecutive in both directions, a selection signal making each processor operate, for each movement, with the corresponding register data   to the selected reference pixel of the current area.   Device according to Claim 12, characterized   in that the correlator operates on a two-dimensional window   the pixel with current blocks subdivided into zones containing 4 pixels each, with a single reference pixel selected by zone, a two-dimensional matrix of registers storing, at a given instant, the pixels of said window according to the same scheduling, four registers being assigned to the processing of a zone, the rows of said registers communicating with each other, in the two directions of the matrix, even-order pixels of the window and other rows of registers communicating with each other order pixels odd of the window, this communication being unidirectional in one direction of the window and being bidirectional in the other direction of the window, a processor being assigned to the processing of each displacement vector and operating with four registers containing even and odd pixels of the window, the order numbers of said pixels taken two by two being consecutive in each direction, a signal of common selection making the processors operate with the data of the registers corresponding to the   reference pixels selected in the current block.